The Self-Play Tutorial

This week, I've been putting together a new project, something that I had in fact considered about this time last year.  I did some preliminary work on the wiki with it, but then lost interest.  I wasn't in the right head space.  But while travelling last week, following the encounter at Butchart Gardens, it felt right to undertake the idea again, and so I have.

On the surface, this is going to seem like an old man's response to the collapse of my campaign, which is fair.  But I think there's more here than that.  For one, some months ago JB and I both discussed the subject of teaching D&D in some depth.  For another, I've been looking for a way to obtain a youtube presence since the middle of Covid.  For another, it wouldn't be my first attempt at building a better dungeon generator.  Although yes, it probably fills up a hole that's just opened, it's not wildly out of character for me to do something like this and honestly, I think that if it can be sustained, there's considerable merit here.

One important change that's come to my aid is, and I'm sorry to bring it up again, chatGPT.  Where it comes to D&D, though, it's a pretty frustrating program to work with ... not because of its design, but because of what it dredges up from it's D&D source material online.  Ask it to invent different entrances and it will gleefully rush to give ten examples of the same result, labeled with alternate "cool"-sounding adventure D&D tropes.  It's a narrow crevice entrance ... no, it's a jumble of rocks entrance.  No, it's a crumbled archway entrance; no, it's a rockfall barrier; no, an overgrown archway; no, a hidden ravine; no, a concealed tree hollow; no, a camouflaged burrow; no, a disguised well entrance ... and so on.  These are all real answers I got when I asked, "Give me different kinds of entrance ways into a dungeon."

It's very difficult to explain, when talking D&D, that these aren't really "different."  Hole in the ground covers all of them.  Granted, most dungeon entrances are a hole in the ground ... but it never occurred to chat that the dungeon, potentially located many score miles into the wilderness, might logically have a palisade or a stone wall.  True enough, a crevice will hide a dungeon from adventurers, if they have to find that crevice in a 10-mile wide hex.  But animals can smell food through a crevice, giant social insects can enter a crevice and plant their eggs there, huge spiders will no doubt find it and make a home there ... at some point, it makes sense to put something in front of that crevice that makes it more secure.  But chat couldn't think of that (I had to), because the entirety of its D&D source material couldn't think of that.

But when I have an idea, and I want to go back and forth on it, filling out details, letting chat take a swing, then writing a new draft, letting chat edit that, whereupon I write a new draft, and chat goes at it again, I can compress a lot of brainstorming and step around troublesome writer blocks with remarkable ease.  Chat can't do the work; it's utterly useless at that.  But it can suggest other words and throw out half-ideas that can be grabbed and run with.  Over the last few days, I've progressed in this particular effort by leaps and bounds because, having thought of dungeon generation for 40+ years, I finally have a real source that I can at least collaborate with.  If the reader will go look at the comments beneath the RDG post I linked above, you'll see that, in fact, in 2014, we were all positing something like what chatGPT can do for me now.

Take this one from Eric:

It'll be a real feat to make a good random dungeon that can be meaningfully revealed one room at a time ... "OK, you beat that pack of goblins, they go down screaming! Next room is..[roll].. a pack of goblin guards ... [roll] ... listening alertly."

That's quintessentially the trouble with every generator: repeat results.  Twice in a row is fine, but it's a disaster when that twice in a row happens repeatedly and gives the message that the goblins are endless; that killing the next group of them produces no real result.

The key, I think, is to control the generation so that it's not just generating from a large random sample.  It's what I'm attempting to do by separating different kinds of dungeon and acknowledging that the path through the "tunnels" dungeon has to be progressive, from space to space ... not treating every space with the same table and the same results.  A tunnel-dungeon is defined by it's rooms, not by its inhabitants; the randomly generated layout has to take that into account, creating just so many of each kind, and in a particular order, with the contents of each room having characteristics that apply to that room.  It's a brutal expectation from a generator-design standpoint, but it's the only real way to get around repetitive results.

My intent is to update my progress every 4,000 words or so.  The page itself, of course, is always going to be in the same place, so if anyone wants to see how it's going, they can.  Meanwhile, with my "collaborator," I'm going to see how far I can take this.

I think it's possible to take it out of the dungeon, though not easily.  I think it's possible to take it into urban areas as well.  I've done some preliminary work for that with my hammers/coins/food pagemaking.  I think it's possible to open up the process to suggestions from outside ... even to adapt what I'm doing to reader interaction regarding what the "player characters" choose to do next.  I have plans for this and youtube, which should reach fruition in a week or two, if I continue not to work on the Streetvendor's Guide, where my head ought to be ... but damn it, this is a new project, and new projects are fun!

Anyway, it's just a start.  We can look at this again when I double it's size so far.

I really like how, at the very end there, I got to write a couple of sentences that actually sounded like playing D&D ...